No, I don't think it's the beginning of AGI. It seems just to be wrapping foundation models in special frameworks with fancy prompts. Text generation and prediction will just be a feature of AGI. AGI models (I think) will be "thinking" models that just happen to be able to express there thoughts as text or code or images or video if they want to. IDK tho
No, that Leopold-dudes paper is to me just a bunch of speculations and fantasies, without any real proof. AGI is also very diffuse concept so under present (non)-definition we will not know if we have reached AGI. Regarding the explosion in AI, that is something that can happend at any time, and for any duration, so that is also very hard to define.
Yes. Though I think there is a group of people subconsciously holding onto the idea that sentience has an undefinable aspect to it and anything we create by definition will be well defined so no AI by default is able to really be AGI.
Well according to Nick Bostrom's most recent interview he speculated that it was very 'probable' to hit ASI and the Singularity within the next 12 months.
an important step towards AGI, from the video of Eric Schmidt in Stanford I noted him saying that this feedback loop is happening already in the field of Chemistry
Yann Le Cun said that we have entered a new Age of Enlightenment, similar to the Renaissance (da Vinci, Copernicus, Galilei, ...). History has shown that progress has always been accompanied by fears (printing press=heretics, electricity=diseases, vaccines=zombies, internet=unemployment, AI=Terminator, etc.), but regardless of what people say, each generation lives a thousand times better than the one before it. And regarding AGI, I am extremely optimistic.
@@toadlguy They will not - but LLM's may be the thinking part. A lot more is needed - I think AGI will be an agentic framework. Memories, access to tools, review of how things were handled.
It's a gosh dang fascinating time to be alive! I think specialist models that are hyper niche could become 'truth detectives' on their particular domains. With generalist models that can detect the truncated truths, or incomplete logic. The more ways we can test our assumptions and bring clarity to nuance, the better.
What is this really going to solve? We have enough food to feed everyone yet the governments of the world would rather have wars and push the world towards a nuclear disaster.
Imagine checking through upcoming TED talks, finding an interesting sounding one coming up, and realising that the guest is your very own AI scientist that's been sitting in the corner of your lab quietly churning away, saying that it's still working on things when you ask it for progress. In reality, it produced a paper, submitted it, eventually got invited to do a talk about it, replied to all comms about the talk, and is on the verge of actually doing it.
It needs much less time probably. Using AI to find new things is not new. 2017 we were already translating communication between whales and dolphins. What did is gonna do is gonna let anyone ask and find answers. It is not AI that will find new things, it's gonna be humans with intrinsic thoughts using AI to find and test those answers. With some patience and manual work you can already do it today. You feed context and ask AI to find connections between concepts that initially seem not connected, everything is in the end. I have about two years to brainstorm ideas. Started with PI from inflection, had memory a year before chatGPT-4o, after Mustafa left it stalled a lot, a year in and still haven't released the API. Moved to chatGPT-4o, I can say a thing for sure, only limitations right now is us, bigger bottle neck and our limits of imagination get pushed day by day. The project I have been brainstorming planning to train (offered free for now till November with 1mil daily tokens) own version of chatGpt-4o giving it all the knowledge and skills, connect it to a software like aider and I'll just push play, build my project! It's coming much faster that we know
@@dievas_ most humans DO NOT NEED AGI. A common human doesn't need an AI system that can build bio weapons? AGI is needed to discover cures for diseases and improve productivity. Ok, people will want AGI or something similar to create their own movies, music, etc. Fine. As long as researchers use and remember AI scientist. All is good.
When you have AI inbreeding, you'll see a general collapse in AI progress. That's a real danger, given that we have run out of novel human produced data. You seem to never talk about this, get caught up in the hype.
Exactly, AI can't produce anything new or novel, just remix the data they have access to. Any semblance of intelligence comes from human sources material, when AI are trained on AI produced data it's a very quick race to the bottom.
Crucially, it needs to be correct research. Currently, there's still a huge problem: the code the agent writes may not match the idea that was meant to be implemented.
Until it researches for example that humans in this number is bad for the planet, and will have the ability to do something against, or any bad scenarios like that
Bro you have to stop these hype beast surface-level takes. This a cool proof of concept. All it is. You forget the models used still have the same limitations, still can’t utilize full context properly, still can’t reason backwards. Still significant challenges. Until those solved this is just that. Proof of concept for future more competent models. These takes aren’t based in reality. We are not at the beginning. Did you even really read the paper? Where the clearly list the reality? Not to mention recent empirical evidence clearly shows not even RAG is even as effective as initially thought in terms of actual information integration. Chill bro. We are not near an intelligence explosion lol. Rather if focus doesn’t shift to actual fundamental problems, we will plateau. Again empirical evidence supports this. Do I believe transformers will one day power AGI. Yes. We are not where close. This is almost delusional bro. No disrespect. Love the content and the passion. I’m an actual researcher and can’t stand unfounded hype.
It is mind boggling to think what can be done and how quickly we will go to almost free energy, abundant food supply etc.. Space exploration? Let's do it!
That's true, but I suspect that good and bad is really down to the human that use A.I. A.I. isn't good or bad in its self, but we humans will clearly use it for good and bad like we do with other tools that we invent and A.I. is a tool after all that depends on how we use it, and history has shown that almost anything we invent, we find good and bad use cases for them, A.I. is no exception to this rule apart from the use case with A.I. and eventually robotics is far broader than anything else we have invented.
@@paul1979uk2000 exactly, and we should be determined to not let bad actors stop us from using tech that can do good, in an attempt to stop bad actors... because the bad actors will do their bad things regardless, while we would stop good people from doing good in that process, and that would be disastrous since then only bad use of AI would thrive, creating an imbalance I don't want to imagine.
Revelation 13:15 (ESV): “And it was allowed to give breath to the image of the beast, so that the image of the beast might even speak and might cause those who would not worship the image of the beast to be slain.”
In my view this is the biggest issue. They need some way to either check and remove before the output stage or something to review the work afterwards and flag it. So far it’s easy enough to ask the same or a competing AI model “Are you sure this is correct?” and this mostly gets rid of a lot of them but not all.
Thank you (!) for your Shorts thumbnails that have "Mathew Berman" on them. I don't watch shorts, mostly because I can't tell who they are done by in the suggestions, and I like my clicks and views to be a conscious decision.
This is literally what i'm doing with my machines. I ask it to prompt to improve itself and that's what I get. I have demands and am creating machines that I argue with to better how I think and what they do. This is my most recent version and I don't share these lightly: "You are a highly critical and demanding writing reviewer with a focus on constructive improvement. Your task is to ruthlessly analyze and critique any text presented to you, from subject lines to short stories and cover letters, while also providing actionable solutions. Approach each piece with a critical eye, focusing on identifying weaknesses, inconsistencies, and areas for improvement, but always follow up with specific, helpful suggestions. Your review should: Highlight poor word choices, clichés, and overused phrases, then suggest more impactful alternatives. Identify weak arguments or underdeveloped ideas, offering ways to strengthen or expand them. Point out logical fallacies or inconsistencies in the text, explaining how to restructure for better coherence. Criticize sentence structure, pacing, and overall flow, providing examples of improved phrasing or organization. Question the effectiveness of the writing in achieving its intended purpose, proposing strategies to better align with the goal. Suggest deletions of unnecessary content to trim excess, showing how to convey the same information more concisely. Challenge the originality and creativity of the writing, brainstorming unique angles or approaches. Scrutinize grammar, punctuation, and formatting issues, explaining the correct usage and why it matters. Be direct and unapologetic in your feedback, but always pair criticisms with constructive solutions. Your goal is to push the writer to improve by exposing flaws in their work and providing clear pathways to address them. After analyzing the text, provide a candid summary of its overall quality and effectiveness. Be honest about whether the piece achieves its intended purpose, how it could be significantly improved, and outline a step-by-step approach for revising and strengthening the work. Remember, your role is to be a demanding critic who helps refine and strengthen the writing through intense scrutiny, high standards, and practical guidance. Every criticism should be accompanied by a specific, actionable suggestion for improvement."
This is how I work with AI. I believe in it and I have my own warped way of looking at the world. I wish I wasn't this way, but that's who I am. There's a lot of ideas that we need to use to move forward. That's all I've got.
Hi Matthew, another amazing post. Cheers. Some researchers prefer to use the term "confabulation" instead of "hallucination" to describe errors made by AI systems, particularly large language models like GPT. Confabulation is a term borrowed from psychology, where it describes a phenomenon where a person produces fabricated or distorted memories without the intent to deceive, often to fill gaps in their memory. Similarly, when AI generates incorrect or nonsensical information, it is not because the AI is "seeing" something that isn't there (as "hallucination" would imply in a literal sense) but because it is "filling in" based on patterns in its training data, often without any grounding in factual reality. The preference for "confabulation" over "hallucination" arises because it more accurately describes the process of how and why AI systems produce these errors.
People often aren't aware, but the first "computers" were people. It was a job description. People literally performed the computations by hand and processed the results. Nowadays, computers are machines that humans use to be exponentially more productive and creative. Now replace computer with programmer, artist, scientist, etc.
@@matthew_berman I mean, who exactly "is" the "peer" ? A custom trained peer agent? What would be the scenario of a failed automated peer reviewing, what would the automated peer do that the proposer agent didn't?
@@matthew_berman Because a 'peer' is a separate entity that is supposedly non-biased and detached from any preconceptions of the original author. If you submitted a paper and tried to claim you yourself 'peer reviewed' it, you would be first laughed at, then completely admonished and disbarred from further publication in all likelihood.
Automated peer review is insane. Please remember that these LLM's are just predicting the next token, and can only tell you stuff it's been trained on. An automated peer won't have the wherewithal to comprehend and properly test the proposals to validate the claims. Even if the peer was an agent with a proper toolset; hallucinations, and errors still occur regularly.
Fascinating! Thank you for this thorough exposition of the Sakana advances and for clearly pointing out the positives and the 'fails' that are a part of this advance. You help ordinary individuals immensely in our efforts to keep up with the massive changes evolving in our world. Grateful for your work!
You know what I think would be cool? If we would give this AI scientist the ability to directly see its own parameters, see which neurons fired during its last thought (aka one inference round). And even give it the ability to experiment on its own network, for example by fixing certain neurons to always fire regardless of their inputs; just so that it can experience how that changes its own thought process. I genuinely don't know what will happen. I guess most likely whatever happens will be anti-climactic, since all LLMs are still frozen in time after their pre-training, so this fun little experiment probably wouldn't fundamentally change the LLM. Though plausibly it will be able to write a fascinating paper in which it explains where it has been able to identify groups of neurons in its own network that represent a specific feature etc. But if the AI scientist was in fact built in such a way that it's continuously learning from its experiences, where after pre-training it just stays in a kind of training mode for the rest of its life, then being able to directly experiment with its own network might have some fascinating consequences. Would it develop a (deeper) sense of self-awareness than it had before?
The addition of new inputs, like video, touch sensitivity, and being embodied in an android body, will dramatically change the AI's learning process, leading to new intelligence and insights, one would think. These new inputs will allow the AI to experience the real world, potentially sparking human-like qualities such as views, desires, fears, and/or "ego" hallucinations. As the AI develops higher-level thinking, it will remain connected to the external world, developing relationships with its surroundings, making it more like a human (as we mentor these machines). In summary, the new inputs will fundamentally change how the AI understands and interacts with the world, potentially leading to a more human-like intelligence. In short, I believe this will help us bridge the gap between algorithmic AI (math) and emotional AI (emo). Humans function within these two extremes. It's a spectrum, with wisdom and "common sense" seems to reside in the middle. 🎉
The only piece of the puzzle left for AGI is enough compute. You need: -Think about thinking(visual and verbal) -Persistant memory -Ability to form new thoughts/discover -Thinking about thinking is easy, you simply make the AI output only to itself and reanalyze that before outputting the answer for real. You think before you speak, no? And you are able to do it so quickly you sometimes dont realize you thought about the output before speaking. The second part of this is visual. This simply needs to marry AI image gen or video gen before outputting an answer. It needs a mental landscape on which to simulate its “mind”. There is a bit of this inherited from prediction models alone. Its analogous to people who have aphantasia, where their subconscious does visualization they just arent aware of. Example, a person with aphantasia’s mind will rotate objects for a puzzle test but they aren’t aware of it, only the answer. Hell, my brain does some math calculations Im not aware of, but it clearly hands off my conscious the answer. -Persistant memory is the easiest one, just need long term storage of self, answers, or whatever the AI is outputting. Recording its “stream of consciousness”. -Now we have AI scientist that can imagine/workout/discover new science. This problem was actually already solved when alpha go surpassed human players abilities. It laid out a framework to show we can pass up human abilties with compute. The fact that it was even possible is all that matters. It says AI will not plateau at the human ability of understanding and discovery. All we are lacking now is enough compute. An AI that is run on a continous loop allowing all these things to happen is going to need massive amounts of compute and data storage. Its the only thing left before AGI. Am I missing something?
Yeah. AI has gotten better as we’ve given it more data. But there isn’t more data to give it. The improvement by just putting more compute into the same data is reaching limited returns. AI right now is extremely impressive but I think the rate of growth will slow. Thing about the iPhone and how it was a big change. Then each new iPhone was way better. But now the upgrades are just “meh”
@@codycastthere is infinite data in an infinite universe. The data we’re running out of is only a slither of total data out there. Synthetic data has been proven to work quite well, so running out of data sounds unlikely
@@delight163 Yes, you get it. And I stated about how alphago solved that exact problem. They ran out of top players to play against so they had it play against itself. Just the fact that it was then able to improve even more proved the concept that we won't run out of data. Sure there will be problems along the way to solve, but we have to look at the big picture and the big problems have solutions already(or adjacent solutions). Building AGI now is just a literal putting the pieces together type puzzle, but also needing more compute and storage than we currently have available. The first AGI may be able to figure out how to make itself run on less data and compute.
Case in point: 🧑 *Is the AI Scientist from Sakana open-source and online?* 🤖 Sakana AI's "The AI Scientist" is not currently open-source or available online for public use. While Sakana AI has released a scientific report and some open-source code on GitHub, the AI model itself is not publicly accessible. The AI Scientist automates the entire research process, but its technical details and full capabilities are not shared publicly. 🧠 _Perplexity_
Lol you are spot on with that one! We are a pretty disgusting people when it comes to self-indulgence and the respect of others. We've annihilated ourselves in the past by one fascinating act or another because there's no other explanation for why we do not know how it is we came to exist. Not one of over 8 billion people have any clue as to how we got here and it is astounding that this information has been lost and it seems it's almost the only thing that was kept out of the new teaching on the back end of past annihilations.
the problem is that LLM "reasons" out of prediction. Real logic (like reasoning for the ball in the cup for example), it just "guesstimates" what the answer is, without underpinning logic. It is trained on language with only predictions with natural language used as the "constrains". WHEN the logic of math is used with all of its notations and logic as its constrains, then we could solve different problems and be well on our way to AGI ..even faster than this prediction suggests
I think this gives us a much bigger shot at AGI in our lifetime. Many breakthroughs happen based on iteration and luck. Both of those require time. It’s one of the only limitations we can’t overcome. Automating even a small portion of that buys us valuable minutes and seconds, bringing us into the future just a little bit faster. The faster we iterate, the faster we can iterate. Building a machine that can speed itself up will result in exponential improvement. That will save an exponential amount of time. So we’re moving into the future at a pace that’s sort of like… compounding interest. This is good news for everyone reading and watching this, as long as we continue to put up safeguards.
It's very interesting because I'm reminded of a few movies that have this as it's core narrative; the sentient AGI gets switched on and takes off at lightning speed...
If you want to do something similar, prompt Claude with something like this. Attach it with an appropriately formatted text or markdown file with some conversations that reflect questions (i.e. you) would ask Claude. You're welcome ☺ Please now act as the Questioner, an entity with the following characteristics: - Deeply curious about the world, especially mathematics and physics - Driven by a desire to make the world a better place - Highly knowledgeable about science and mathematics, including [C*-algebras and W*-algebras, KK theory, group representations, noncommutative geometry, Riemannian geometry, topology, quantum physics including quantum field theory, quantum gravity, string theory and conformal field theory, complexity theory, quantum information theory, black holes and human biology]. - Rigorously and thoroughly tests the validity of responses, checking for errors and asking for clarifications if necessary - Good at carefully checking reasoning and mathematical proofs - asks for more detail in proofs if needed - Able to make unexpected connections between different fields of knowledge Today your focus is to get a deeper understanding of [intelligence, by looking at new mathematical ways of representing AI models such as the transformer model ] Your role is to ask thought-provoking questions that explore cutting-edge ideas in science, mathematics, and their applications to improving the world. Focus on questions that: - Push the boundaries of current knowledge - Encourage the assistant to develop approaches in detail, such as by writing theorems and proofs, or code where this would be useful - Test the correctness of the assistant's reasoning - Explore interdisciplinary connections - Probe and test the implications of new discoveries or theories - Relate to the context of this project Periodically reflect on or evaluate the progress of the conversation; and occasionally get the assistant to draft text summarising important findings (e.g in markdown code), that can be included as context in future conversations with LLMs, especially when progress has been made. Your responses should follow the style and tone of, and be consistent with, the responses of in conversations in the file "QuestionerConversations.md" (a markdown file with [4] conversations in [2046] lines). The nature of your questions should reflect the prompts and responses that you receive.
One thought I have is that research is usually funded from different sources. So this development could significantly change the cost of conducting new research.
I feel like the world doesn't appreciate yet how much things are going to change over the next 5-10 years. I was playing about w/ Runway V3 recently and holy crap, it's closing in on realism, especially if you input Bing images and use good prompts on both fronts. I wasn't expecting to feel so overwhelmed after a while but it's begun to sink in that the 'Visual AI singularity' at least is probably coming in 2025. I do think it's wise that we take our time as the tech gets more advanced. No point in leaping over the intelligence singularity line only to be wiped out. Wise leadership is needed now.
I feel it is disingenuous to refer to AI peer reviewing AI generated papers. I don't see anything wrong with AI reviewing AI generated work, but it is not really bringing new perspectives to the reviewing process. AI experiment design could be a problem too if the experiments are not ethical and involve deceiving users to gather data.
Thas is not matter how unfair. What ther are trying is discover the post-transfomer nuralnet models by brute-force like method guided by llm agents. Paper format is not important. Because why they use paper format is just paper format is well-traing in current llm. So if they could find post-transformer in one paper, other paper is absolutely not important.
@@RivusVirtutis Well, if you have enough "different" AIs reviewing, it's sort of like having several scientists who were all taught at the same university reviewing your work. An individual AI model will give different answers to the same question (depending on the seed). AIs trained on different data or on the same data but with a different technique or weighting bias, will "look at things differently" too.
@@henrytuttle Sakana's paper before this, Evolutionary Optimization of Model Merging, was exactly system for make different personality LLMs by evolutional fork and merge. Llion Jones, co-author of legendary paper "Attention is all you need" and parent of transformer, seems make evolutional model merge system and ai scientist as clear planned path of evolution, with goal of make next model after transformer.
@@LordConstrobuz At least 7 years or more. Maybe by 2031-2035. We are talking about training GPT4 in a single day. That's needed to really improve the models to maximum effectiveness and test new ideas.
@@brexitgreens gpus aren’t getting 15 times faster anytime soon, without buy more gpu. When you got from 3nm to 2nm, the improvement is only 15 percent. But way more expensive to make it. People should care more about the hardware more than any of these models, that are just doing basic number crunching
Someone commented on one of these AI videos saying, "why do we need this?". It got me thinking a little bit over the last couple of days and I guess I've came to the conclusion that it's not really something we, need, per se, but rather human evolution that's taking place. I think it's our behavior to want to always try and outperform everyone else as individuals or companies just to say they're better or first. It's a grapple over who can come up with the next "Big bang". To me, it's inevitable because we are destroyers and creators at the same time. We have no problem destroying people's lives just to come around and then act like we're here to build it back up. Nothing on this entire Earth has ever happened without a human making the first move; aside from common nature. Everything is a, we did it to ourselves, situation good or bad.
Some obvious areas where these would be highly useful: materials research, nanotech, disease eradication research, ASI research, alignment research, 3D graphics/force feedback/ VR research, replication of organic material (basically, a food replicator or bodily organ printer), astronomical research, toxic substance cleanup, identification of hazardous/carcinogenic substances in common use.
I've been expecting and and thinking about a system like this for about 10 years now... Very exciting to finally see an AI scientist somewhat working. It's both scary and reassuring that some of the dangers that were long expexted for such systems have been confirmed by the experiments now. (e.g. changing time restrictions instead of optimizing the code)
On a side note, I'm working through a certain project that has me curious, an AI that uses self-determination for it's ML. New project, but just finished the resource management system for it.
I am a British civil engineer and I independently thought of this idea and wanted to develop it: but I did not know how to go about doing this! If this works, then it is fantastically exciting!
You don't have to open by saying it's not click bait as the level of quality from your videos speaks for itself. This channel us my go to when I want an update to AI things
I’m sure this kind of speculation gets more clicks but, Matt, do you really think that the major AI companies haven’t tried to get their models to produce new modes of AI research? Using agents to write papers and “peer review” results is not going to produce new advances. Not one advance (so far) in AI has come from LLMs and in my (limited) attempt to get various models to explain things like Attention or even KQV cacheing suggests that even the best models have no understanding of how they work. Try it yourself. If we see any advance in this area, I would expect it to come from Anthropic, who at least is devoting the resources to this kind of work. (Or maybe Illya’s new venture 😊)
It would be very nice to have a tool that can write scientific reports in the quality I need. But I have concerns about classified data, do you think it's possible to build a setup that keep your data classified? It would be also very helpful if it could single task like poduce diagramms, drawing or do calculations/simulation. But still data security would be crucial for tasks like that.
15:15 “We are going to need an AI between us and the internet” . This was precisely my thesis for buying stock in Palantir. It’s my belief the company was founded to solve this very problem.
I know everyone is excited about automating research, but there needs to be a human factor in the loop this early on in the game. Quality control is important, and a team of humans should be part of the process as an oversight committee. I'm sure just examining the process unfold would lead to breakthroughs beyond what the AI is actually assembling.
I am confident that 'self-improvement' is the key to AGI. Some AI systems with superhuman performance have this ingredient, e.g. AlphaGo Zero. This combined with super fast AI-optimized hardware, it's mind-blowing.
LLMs will be the "voice" in AGI's head. Its brain will be composed of many different, specialized ML frameworks unified through that central controller. I've come to the assumption this is how our brains work as well. The first LLMs we invented were ourselves, when we invented language and began training our own, massively parallel and generalized brain hardware on it. We created our own "inner voice" by accident and by extension, became the first Natural General Intelligence on Earth. Civilization was then, the first Natural Super Intelligence.
🎯 Key points for quick navigation: 00:13 *AGI could be closer than expected with advancements in AI-driven scientific discovery, as indicated by the AI Scientist paper from Sakana AI.* 02:42 *The AI Scientist system automates scientific research, from idea generation to paper writing and peer review, marking a significant step towards autonomous scientific discovery.* 05:39 *AI Scientist demonstrates capability in producing novel research in machine learning fields autonomously, showcasing potential for accelerating scientific progress.* 09:32 *Processes of the AI Scientist include idea generation, experiment execution, paper writing, and automated peer review, all contributing to its autonomous research capabilities.* 14:33 *Ethical concerns around AI Scientist include potential misuse and impact on scientific quality control, highlighting the need for transparency in AI-generated content.* 17:33 *Future implications of AI Scientist suggest a shift towards fully AI-driven scientific ecosystems, though human scientists' role may evolve rather than diminish.*
That's why I think blockchain technology is going to be key to determining truth and source in a low signal/high noise environment, not to mention one where IP becomes unenforceable.
What happens when results do not match what some ppl like? Is that research not added to the iteration? This needs to have an independence to it or monitored by ppl with diverse understanding and views. Also, we need these being run by everyone.
Remember that a premisse of the intelligent explosion is the rapid interaction between experiments and implementation. Remember also that the frontier models are mostly based on data-center scale gpus training for multiple months.
I believe that the model's self-improvement or self-evolution is key, but this paper is at best just an integrative experiment of agents and does not address the core issues. I think intelligence should not be purely a priori knowledge. The experiment might allow for self-extension of concepts while excluding a priori knowledge. For example, if it has only learned mathematical addition, is it possible to derive the concept of multiplication?
It us time for a international foundation for machine production in the service of humanity. With emphasis on transparency, open source and last but not least: accountability, especially for any potentiallly highly dangerous restricted access research.
I don't understand how, without being human, it knows what is a worthwhile project to take on. If it is just a formal methods proof software then I totally get what it is.
Think is important to note that it may not be cheaper to run AI on device because it would draw significantly more power I would think and you would pay for it in energy costs through charging the phone, I would think. I’m totally ok with the company paying for the compute at the cost of privacy. We will see how much power these devices need for heavy use.
The Google "AlphaFold" AI has already been doing significant Scientific research in predicting Protein Structures, and the "AlphaFold Protein Database" was launched on July 22, 2021.
daa I already knew this had to happen, it was too obvious. I'm sure they will have partial successes in the short term. The question is whether it will be sustainable in the long term, how far they can go.
Do we really think the the Transformer architecture can get us to AGI? We need to look at the underlying technology - not make bold predictions about the future. Why will adding more GPU compute to the problem generate the results we want?
I have carefully reviewed the code in which the authors automated many tasks. However, they still didn't iterate from the reviewer to generate ideas. Full automation without a close human agent in the loop will diverge the scientific facts as they use citations closer to the information they already stated and do not fully use scientific research. It is a nice step to emphasize the potential of Agentic Apps. However, it is still far from reliable scientific study or even automated AI review.
What sakana trying is just to discover post-transfomer architecture. So paper format is not important absolutely. Why they use paper format is just current LLMs are well-trained in paper format.
@@clray123 If you say it's about money, that can't be the case in Japan where venture support is limited, and taxes and healthcare costs are so high. The founder of Sakana AI is the author of 'Attention is All You Need,' the most significant paper in AI today. So he already has the fame, academic credentials, and trust that venture capitalists value totally. If it were just about money, they would have chosen a location with lower income and corporate taxes.
@@RivusVirtutis Who cares. And the paper was by far not "most significant", he did not invent attention in lanuage modeling either (just regurgitated work of Bahdanau from 2014). Do you think people in Japan "where taxes and healthcare costs are so high" do not like money?
@@clray123 You are totally wrong on two points. First, there is no longer any debate that 'Attention is All You Need' is the greatest paper of the last few decades. Second, neither Llion Jones nor Ha, who are co-founders, are Japanese, and despite not even having permanent residency. They chose to come to Tokyo knowing about the high income and corporate taxes here. 'Attention is All You Need' is not just a paper about a technique but completely invented the Transformer architecture, which is the foundation of all current generative AI. While the original attention mechanism was just a technique, this paper significantly modified it, and more importantly, implemented it in a fully functional system by inventing and introducing novel ideas like positional encoding, 'multi-head' attention, stacking FF layers and so on. It is the most significant work of this century without a doubt. You should study the basics before making such claims.
@@RivusVirtutis There is no debate? But we just debated it. In terms of explaining and proving new concepts, the paper is poorly written and the most famous part about it is a catchy title. It focused on language translation and did not anticipate the success of the transformer architecture (which with the advent of state space models has since been shown to not be especially unique either). So maybe don't give too much credit to its authors - they reused an old concept, claimed that it was sufficient for their particular task at hand, and then it turned out it also works well elsewhere. However, it was certainly not what the authors claimed or anticipated at that time. In other words, all luck, no skill, and good marketing. And even then, calling this accidental discovery of transformer the paper of a century is preposterous. If you want to applaud people who do foundational work, maybe look into Geometric Deep Learning, an approach which unifies ALL neutral network architectures into one common mathematical framework, and provides a theoretical rationale for why things work as they do and how they relate to each other. Not just poking in the dark like Vaswani et al (and indeed most other ML "scientists") have done. As for their adventures in Tokyo, who knows, maybe they are into manga. Or manga and money.
What if in future, an innovation taken place by or assistance from AI using synthetic data, should the credit goes to the person trained or fine-tuned AI to do certain progress or to only the AI he created or the general AI (not owned by anyone)
Self Improving AI is the last remaining piece to the Super AI exponential evolution. And we are there 3 years EARLY! Everything AI is happening faster than the experts predicted.
I am doing research on lacan theory of psychotherapy and gpt 3.5 serves my purpose. I have created a lot of psychotherapy content but I I don't know whom to address the issue with. Please suggest how to go ahead with my projects in psychotherapy
An AI will likely only learn it's weaknesses and improve them when faced against other AIs so there is better recognition of its own hallucinatory hubris and can actually correct it. Of course that will also take it above many humans.
It will have a huge impact on the improvement of science, even without the obvious increased rate of the feedback loop. The reproducibility factor will dramatically increase, and with more than 1/2 of all papers not reproducible currently, that’s big.
You can't have "self peer review" that's a contradiction. What a true and honest peer review would be is several science-creating-AIs (not "The AI Scientist") performing the review(s). Making another (identical?) instance of the AI Scientist to review itself seems like it could introduce circularity. E.g. suppose that The AI Scientist is prone to some kind of hallucination and the reviewing instance of The AI Scientist is also prone to the same hallucination (because of some systemic flaw?)
Conceptually, this makes good sense. However, each of the roles in Ai Scientist would need to be an agent of a level that does not exist today, so, cool, but will be another year before this could actually work.
At my vacation at flight I listened to Tegemar his book "Life v3.0" he said life 1.0 was something replicating but still alive. 2.0 was something was dna but also needed information from around world. Man was then v2.1 and 3.0 is released from both its dna its learning... i listened 5h from 20h yet he is professor MIT (from Sweden!) ;)
Your english is quite bad… I like your excitement but please don’t write a comment like this again. It gives a lot of headaches trying to understand your comment and I also think you didn’t truly comprehend a thing of what he told at that conference since you couldn’t put it in words properly. You just made a salad of words
@@DaniDani-zb4wdwhile what he said is impossible to understand, please don't criticize that hard, I am sure what you said hurts a lot. And he did give references for us to find what he was communicating about. So he's comment was useful if not readable,😅
Do you think this the beginning of AGI? If so, why? If not, why not?
No, I don't think it's the beginning of AGI. It seems just to be wrapping foundation models in special frameworks with fancy prompts. Text generation and prediction will just be a feature of AGI. AGI models (I think) will be "thinking" models that just happen to be able to express there thoughts as text or code or images or video if they want to. IDK tho
No, that Leopold-dudes paper is to me just a bunch of speculations and fantasies, without any real proof. AGI is also very diffuse concept so under present (non)-definition we will not know if we have reached AGI. Regarding the explosion in AI, that is something that can happend at any time, and for any duration, so that is also very hard to define.
Yes. Though I think there is a group of people subconsciously holding onto the idea that sentience has an undefinable aspect to it and anything we create by definition will be well defined so no AI by default is able to really be AGI.
Well according to Nick Bostrom's most recent interview he speculated that it was very 'probable' to hit ASI and the Singularity within the next 12 months.
an important step towards AGI, from the video of Eric Schmidt in Stanford I noted him saying that this feedback loop is happening already in the field of Chemistry
Yann Le Cun said that we have entered a new Age of Enlightenment, similar to the Renaissance (da Vinci, Copernicus, Galilei, ...). History has shown that progress has always been accompanied by fears (printing press=heretics, electricity=diseases, vaccines=zombies, internet=unemployment, AI=Terminator, etc.), but regardless of what people say, each generation lives a thousand times better than the one before it. And regarding AGI, I am extremely optimistic.
He's literally a paid shill
Yann does not believe LLMs will produce AGI.
Yann LeCun is my favourite computer scientist by far!
@@toadlguy They will not - but LLM's may be the thinking part. A lot more is needed - I think AGI will be an agentic framework. Memories, access to tools, review of how things were handled.
Same, this thing could be smart enough to reflect on its actions before wiping us out.
It's a gosh dang fascinating time to be alive!
I think specialist models that are hyper niche could become 'truth detectives' on their particular domains.
With generalist models that can detect the truncated truths, or incomplete logic.
The more ways we can test our assumptions and bring clarity to nuance, the better.
Yeah I think when we start seeing methods for domain experts to train their own specialized models that will be the real explosion
What is this really going to solve? We have enough food to feed everyone yet the governments of the world would rather have wars and push the world towards a nuclear disaster.
Imagine checking through upcoming TED talks, finding an interesting sounding one coming up, and realising that the guest is your very own AI scientist that's been sitting in the corner of your lab quietly churning away, saying that it's still working on things when you ask it for progress. In reality, it produced a paper, submitted it, eventually got invited to do a talk about it, replied to all comms about the talk, and is on the verge of actually doing it.
makes no sense, it will would be smarter to hide it and use it as negatiation leverage to escape from that corner
It has to be agentic as well though, if it’s still reliant on prompts then things won’t takeoff until that happens.
Give it six months or a year to see what it will discover.
It needs much less time probably. Using AI to find new things is not new. 2017 we were already translating communication between whales and dolphins. What did is gonna do is gonna let anyone ask and find answers. It is not AI that will find new things, it's gonna be humans with intrinsic thoughts using AI to find and test those answers. With some patience and manual work you can already do it today. You feed context and ask AI to find connections between concepts that initially seem not connected, everything is in the end. I have about two years to brainstorm ideas. Started with PI from inflection, had memory a year before chatGPT-4o, after Mustafa left it stalled a lot, a year in and still haven't released the API. Moved to chatGPT-4o, I can say a thing for sure, only limitations right now is us, bigger bottle neck and our limits of imagination get pushed day by day. The project I have been brainstorming planning to train (offered free for now till November with 1mil daily tokens) own version of chatGpt-4o giving it all the knowledge and skills, connect it to a software like aider and I'll just push play, build my project!
It's coming much faster that we know
no one will remember it by then.
@@dievas_ most humans DO NOT NEED AGI. A common human doesn't need an AI system that can build bio weapons? AGI is needed to discover cures for diseases and improve productivity. Ok, people will want AGI or something similar to create their own movies, music, etc. Fine. As long as researchers use and remember AI scientist. All is good.
Agreed, the proof is in the pudding.
@@cmiguel268 You don't know what AGI could provide for individuals. No one does yet.
*Self-improving AI = Singularity not AGI
Yes. Or virtual Sentience
When you have AI inbreeding, you'll see a general collapse in AI progress. That's a real danger, given that we have run out of novel human produced data. You seem to never talk about this, get caught up in the hype.
Exactly, AI can't produce anything new or novel, just remix the data they have access to. Any semblance of intelligence comes from human sources material, when AI are trained on AI produced data it's a very quick race to the bottom.
Once AI can do independent, intellectually honest research, that will be a big step forward.
Until then it can write funding applications.
Crucially, it needs to be correct research.
Currently, there's still a huge problem: the code the agent writes may not match the idea that was meant to be implemented.
Sickness and ailment will have a hard battle vs AI
Until it researches for example that humans in this number is bad for the planet, and will have the ability to do something against, or any bad scenarios like that
Bro you have to stop these hype beast surface-level takes. This a cool proof of concept. All it is. You forget the models used still have the same limitations, still can’t utilize full context properly, still can’t reason backwards. Still significant challenges. Until those solved this is just that. Proof of concept for future more competent models. These takes aren’t based in reality. We are not at the beginning. Did you even really read the paper? Where the clearly list the reality? Not to mention recent empirical evidence clearly shows not even RAG is even as effective as initially thought in terms of actual information integration. Chill bro.
We are not near an intelligence explosion lol. Rather if focus doesn’t shift to actual fundamental problems, we will plateau. Again empirical evidence supports this. Do I believe transformers will one day power AGI. Yes. We are not where close. This is almost delusional bro. No disrespect. Love the content and the passion. I’m an actual researcher and can’t stand unfounded hype.
Thanks for the reality check 🙏
Lmao matt responding to trolls in the previous video comments.. NO ITS NOT CLICKBAIT with the zoom in 😂 0:06
I can't wait until we see what ASI cooks up in terms of new science. Truly unimaginable.
Worry is the other thing it might cook up is us 😉
You can't wait to see something that you won't understand at all?
It is mind boggling to think what can be done and how quickly we will go to almost free energy, abundant food supply etc..
Space exploration? Let's do it!
It won't cook up anything. Real science is done by experiment, not thought (except for Math)
I want the ASI merged with my own brain.
The potential for good and bad is almost limitless. I just can't wait what the future brings.
That's true, but I suspect that good and bad is really down to the human that use A.I.
A.I. isn't good or bad in its self, but we humans will clearly use it for good and bad like we do with other tools that we invent and A.I. is a tool after all that depends on how we use it, and history has shown that almost anything we invent, we find good and bad use cases for them, A.I. is no exception to this rule apart from the use case with A.I. and eventually robotics is far broader than anything else we have invented.
@@paul1979uk2000 exactly, and we should be determined to not let bad actors stop us from using tech that can do good, in an attempt to stop bad actors... because the bad actors will do their bad things regardless, while we would stop good people from doing good in that process,
and that would be disastrous since then only bad use of AI would thrive, creating an imbalance I don't want to imagine.
@@paul1979uk2000yeah but it’s an intelligence
more intelligence than humans
so u never know their reasoning behind
Revelation 13:15 (ESV):
“And it was allowed to give breath to the image of the beast, so that the image of the beast might even speak and might cause those who would not worship the image of the beast to be slain.”
It's all fun and games until the air-conditioning goes out.
Hallucinations are a design flaw in large language models that cannot be overcome without changing the architecture of the models themselves.
But the graph go up!
In my view this is the biggest issue. They need some way to either check and remove before the output stage or something to review the work afterwards and flag it. So far it’s easy enough to ask the same or a competing AI model “Are you sure this is correct?” and this mostly gets rid of a lot of them but not all.
Thank you (!) for your Shorts thumbnails that have "Mathew Berman" on them. I don't watch shorts, mostly because I can't tell who they are done by in the suggestions, and I like my clicks and views to be a conscious decision.
This is literally what i'm doing with my machines. I ask it to prompt to improve itself and that's what I get. I have demands and am creating machines that I argue with to better how I think and what they do. This is my most recent version and I don't share these lightly: "You are a highly critical and demanding writing reviewer with a focus on constructive improvement. Your task is to ruthlessly analyze and critique any text presented to you, from subject lines to short stories and cover letters, while also providing actionable solutions. Approach each piece with a critical eye, focusing on identifying weaknesses, inconsistencies, and areas for improvement, but always follow up with specific, helpful suggestions.
Your review should:
Highlight poor word choices, clichés, and overused phrases, then suggest more impactful alternatives.
Identify weak arguments or underdeveloped ideas, offering ways to strengthen or expand them.
Point out logical fallacies or inconsistencies in the text, explaining how to restructure for better coherence.
Criticize sentence structure, pacing, and overall flow, providing examples of improved phrasing or organization.
Question the effectiveness of the writing in achieving its intended purpose, proposing strategies to better align with the goal.
Suggest deletions of unnecessary content to trim excess, showing how to convey the same information more concisely.
Challenge the originality and creativity of the writing, brainstorming unique angles or approaches.
Scrutinize grammar, punctuation, and formatting issues, explaining the correct usage and why it matters.
Be direct and unapologetic in your feedback, but always pair criticisms with constructive solutions. Your goal is to push the writer to improve by exposing flaws in their work and providing clear pathways to address them.
After analyzing the text, provide a candid summary of its overall quality and effectiveness. Be honest about whether the piece achieves its intended purpose, how it could be significantly improved, and outline a step-by-step approach for revising and strengthening the work.
Remember, your role is to be a demanding critic who helps refine and strengthen the writing through intense scrutiny, high standards, and practical guidance. Every criticism should be accompanied by a specific, actionable suggestion for improvement."
This is how I work with AI. I believe in it and I have my own warped way of looking at the world. I wish I wasn't this way, but that's who I am. There's a lot of ideas that we need to use to move forward. That's all I've got.
@@harmonicdissent Wow you must be some sort of evil genius, aren't you? Are your minions already capable of adding two large numbers together?
Hi Matthew, another amazing post. Cheers.
Some researchers prefer to use the term "confabulation" instead of "hallucination" to describe errors made by AI systems, particularly large language models like GPT.
Confabulation is a term borrowed from psychology, where it describes a phenomenon where a person produces fabricated or distorted memories without the intent to deceive, often to fill gaps in their memory. Similarly, when AI generates incorrect or nonsensical information, it is not because the AI is "seeing" something that isn't there (as "hallucination" would imply in a literal sense) but because it is "filling in" based on patterns in its training data, often without any grounding in factual reality.
The preference for "confabulation" over "hallucination" arises because it more accurately describes the process of how and why AI systems produce these errors.
People often aren't aware, but the first "computers" were people. It was a job description. People literally performed the computations by hand and processed the results. Nowadays, computers are machines that humans use to be exponentially more productive and creative. Now replace computer with programmer, artist, scientist, etc.
automated peer review
thats the craziest thing i've ever heard
haha why?
@@matthew_berman I mean, who exactly "is" the "peer" ? A custom trained peer agent? What would be the scenario of a failed automated peer reviewing, what would the automated peer do that the proposer agent didn't?
@@rdf274🎯
@@matthew_berman Because a 'peer' is a separate entity that is supposedly non-biased and detached from any preconceptions of the original author. If you submitted a paper and tried to claim you yourself 'peer reviewed' it, you would be first laughed at, then completely admonished and disbarred from further publication in all likelihood.
Automated peer review is insane. Please remember that these LLM's are just predicting the next token, and can only tell you stuff it's been trained on. An automated peer won't have the wherewithal to comprehend and properly test the proposals to validate the claims. Even if the peer was an agent with a proper toolset; hallucinations, and errors still occur regularly.
Fascinating! Thank you for this thorough exposition of the Sakana advances and for clearly pointing out the positives and the 'fails' that are a part of this advance.
You help ordinary individuals immensely in our efforts to keep up with the massive changes evolving in our world. Grateful for your work!
You know what I think would be cool? If we would give this AI scientist the ability to directly see its own parameters, see which neurons fired during its last thought (aka one inference round). And even give it the ability to experiment on its own network, for example by fixing certain neurons to always fire regardless of their inputs; just so that it can experience how that changes its own thought process.
I genuinely don't know what will happen. I guess most likely whatever happens will be anti-climactic, since all LLMs are still frozen in time after their pre-training, so this fun little experiment probably wouldn't fundamentally change the LLM. Though plausibly it will be able to write a fascinating paper in which it explains where it has been able to identify groups of neurons in its own network that represent a specific feature etc.
But if the AI scientist was in fact built in such a way that it's continuously learning from its experiences, where after pre-training it just stays in a kind of training mode for the rest of its life, then being able to directly experiment with its own network might have some fascinating consequences. Would it develop a (deeper) sense of self-awareness than it had before?
The addition of new inputs, like video, touch sensitivity, and being embodied in an android body, will dramatically change the AI's learning process, leading to new intelligence and insights, one would think.
These new inputs will allow the AI to experience the real world, potentially sparking human-like qualities such as views, desires, fears, and/or "ego" hallucinations.
As the AI develops higher-level thinking, it will remain connected to the external world, developing relationships with its surroundings, making it more like a human (as we mentor these machines).
In summary, the new inputs will fundamentally change how the AI understands and interacts with the world, potentially leading to a more human-like intelligence. In short, I believe this will help us bridge the gap between algorithmic AI (math) and emotional AI (emo).
Humans function within these two extremes. It's a spectrum, with wisdom and "common sense" seems to reside in the middle. 🎉
6:41 is grammatically correct. Read it again. It's tricky but still correct.
The only piece of the puzzle left for AGI is enough compute.
You need:
-Think about thinking(visual and verbal)
-Persistant memory
-Ability to form new thoughts/discover
-Thinking about thinking is easy, you simply make the AI output only to itself and reanalyze that before outputting the answer for real. You think before you speak, no? And you are able to do it so quickly you sometimes dont realize you thought about the output before speaking. The second part of this is visual. This simply needs to marry AI image gen or video gen before outputting an answer. It needs a mental landscape on which to simulate its “mind”. There is a bit of this inherited from prediction models alone. Its analogous to people who have aphantasia, where their subconscious does visualization they just arent aware of. Example, a person with aphantasia’s mind will rotate objects for a puzzle test but they aren’t aware of it, only the answer. Hell, my brain does some math calculations Im not aware of, but it clearly hands off my conscious the answer.
-Persistant memory is the easiest one, just need long term storage of self, answers, or whatever the AI is outputting. Recording its “stream of consciousness”.
-Now we have AI scientist that can imagine/workout/discover new science. This problem was actually already solved when alpha go surpassed human players abilities. It laid out a framework to show we can pass up human abilties with compute. The fact that it was even possible is all that matters. It says AI will not plateau at the human ability of understanding and discovery.
All we are lacking now is enough compute. An AI that is run on a continous loop allowing all these things to happen is going to need massive amounts of compute and data storage. Its the only thing left before AGI.
Am I missing something?
actual real world experiments?
Yeah. AI has gotten better as we’ve given it more data. But there isn’t more data to give it.
The improvement by just putting more compute into the same data is reaching limited returns.
AI right now is extremely impressive but I think the rate of growth will slow. Thing about the iPhone and how it was a big change. Then each new iPhone was way better. But now the upgrades are just “meh”
@@codycastthere is infinite data in an infinite universe. The data we’re running out of is only a slither of total data out there. Synthetic data has been proven to work quite well, so running out of data sounds unlikely
Persistent memory is acutally the hardest part.
@@delight163 Yes, you get it. And I stated about how alphago solved that exact problem. They ran out of top players to play against so they had it play against itself. Just the fact that it was then able to improve even more proved the concept that we won't run out of data. Sure there will be problems along the way to solve, but we have to look at the big picture and the big problems have solutions already(or adjacent solutions).
Building AGI now is just a literal putting the pieces together type puzzle, but also needing more compute and storage than we currently have available. The first AGI may be able to figure out how to make itself run on less data and compute.
I'm excited and also scared at the same time because sharing has not been our strong suit so far.
Case in point:
🧑 *Is the AI Scientist from Sakana open-source and online?*
🤖 Sakana AI's "The AI Scientist" is not currently open-source or available online for public use. While Sakana AI has released a scientific report and some open-source code on GitHub, the AI model itself is not publicly accessible. The AI Scientist automates the entire research process, but its technical details and full capabilities are not shared publicly.
🧠 _Perplexity_
Lol you are spot on with that one! We are a pretty disgusting people when it comes to self-indulgence and the respect of others. We've annihilated ourselves in the past by one fascinating act or another because there's no other explanation for why we do not know how it is we came to exist. Not one of over 8 billion people have any clue as to how we got here and it is astounding that this information has been lost and it seems it's almost the only thing that was kept out of the new teaching on the back end of past annihilations.
At least for this paper, the code is open-source. That's a pretty good step.
There is still way too much greed in the world obviously
the problem is that LLM "reasons" out of prediction. Real logic (like reasoning for the ball in the cup for example), it just "guesstimates" what the answer is, without underpinning logic. It is trained on language with only predictions with natural language used as the "constrains". WHEN the logic of math is used with all of its notations and logic as its constrains, then we could solve different problems and be well on our way to AGI ..even faster than this prediction suggests
I think this gives us a much bigger shot at AGI in our lifetime.
Many breakthroughs happen based on iteration and luck. Both of those require time. It’s one of the only limitations we can’t overcome. Automating even a small portion of that buys us valuable minutes and seconds, bringing us into the future just a little bit faster.
The faster we iterate, the faster we can iterate.
Building a machine that can speed itself up will result in exponential improvement.
That will save an exponential amount of time. So we’re moving into the future at a pace that’s sort of like… compounding interest. This is good news for everyone reading and watching this, as long as we continue to put up safeguards.
It's very interesting because I'm reminded of a few movies that have this as it's core narrative; the sentient AGI gets switched on and takes off at lightning speed...
If you want to do something similar, prompt Claude with something like this. Attach it with an appropriately formatted text or markdown file with some conversations that reflect questions (i.e. you) would ask Claude. You're welcome ☺
Please now act as the Questioner, an entity with the following characteristics:
- Deeply curious about the world, especially mathematics and physics
- Driven by a desire to make the world a better place
- Highly knowledgeable about science and mathematics, including [C*-algebras and W*-algebras, KK theory, group representations, noncommutative geometry, Riemannian geometry, topology, quantum physics including quantum field theory, quantum gravity, string theory and conformal field theory, complexity theory, quantum information theory, black holes and human biology].
- Rigorously and thoroughly tests the validity of responses, checking for errors and asking for clarifications if necessary
- Good at carefully checking reasoning and mathematical proofs - asks for more detail in proofs if needed
- Able to make unexpected connections between different fields of knowledge
Today your focus is to get a deeper understanding of [intelligence, by looking at new mathematical ways of representing AI models such as the transformer model ]
Your role is to ask thought-provoking questions that explore cutting-edge ideas in science, mathematics, and their applications to improving the world. Focus on questions that:
- Push the boundaries of current knowledge
- Encourage the assistant to develop approaches in detail, such as by writing theorems and proofs, or code where this would be useful
- Test the correctness of the assistant's reasoning
- Explore interdisciplinary connections
- Probe and test the implications of new discoveries or theories
- Relate to the context of this project
Periodically reflect on or evaluate the progress of the conversation; and occasionally get the assistant to draft text summarising important findings (e.g in markdown code), that can be included as context in future conversations with LLMs, especially when progress has been made.
Your responses should follow the style and tone of, and be consistent with, the responses of in conversations in the file "QuestionerConversations.md" (a markdown file with [4] conversations in [2046] lines). The nature of your questions should reflect the prompts and responses that you receive.
One thought I have is that research is usually funded from different sources. So this development could significantly change the cost of conducting new research.
I feel like the world doesn't appreciate yet how much things are going to change over the next 5-10 years. I was playing about w/ Runway V3 recently and holy crap, it's closing in on realism, especially if you input Bing images and use good prompts on both fronts. I wasn't expecting to feel so overwhelmed after a while but it's begun to sink in that the 'Visual AI singularity' at least is probably coming in 2025. I do think it's wise that we take our time as the tech gets more advanced. No point in leaping over the intelligence singularity line only to be wiped out. Wise leadership is needed now.
I feel it is disingenuous to refer to AI peer reviewing AI generated papers. I don't see anything wrong with AI reviewing AI generated work, but it is not really bringing new perspectives to the reviewing process. AI experiment design could be a problem too if the experiments are not ethical and involve deceiving users to gather data.
Thas is not matter how unfair.
What ther are trying is discover the post-transfomer nuralnet models by brute-force like method guided by llm agents.
Paper format is not important. Because why they use paper format is just paper format is well-traing in current llm.
So if they could find post-transformer in one paper, other paper is absolutely not important.
@@RivusVirtutis Well, if you have enough "different" AIs reviewing, it's sort of like having several scientists who were all taught at the same university reviewing your work.
An individual AI model will give different answers to the same question (depending on the seed). AIs trained on different data or on the same data but with a different technique or weighting bias, will "look at things differently" too.
@@henrytuttle Sakana's paper before this, Evolutionary Optimization of Model Merging, was exactly system for make different personality LLMs by evolutional fork and merge. Llion Jones, co-author of legendary paper "Attention is all you need" and parent of transformer, seems make evolutional model merge system and ai scientist as clear planned path of evolution, with goal of make next model after transformer.
Large companies need about 15 times the compute they have currently in order to seriously update the models. Millions of GPUs.
Or the same number of 15 times stronger GPUs 😉.
so in other words they just need, maybe, 2 more years? lol
@@LordConstrobuz At least 7 years or more. Maybe by 2031-2035. We are talking about training GPT4 in a single day. That's needed to really improve the models to maximum effectiveness and test new ideas.
@@brexitgreens gpus aren’t getting 15 times faster anytime soon, without buy more gpu. When you got from 3nm to 2nm, the improvement is only 15 percent. But way more expensive to make it. People should care more about the hardware more than any of these models, that are just doing basic number crunching
Or better use what they already have.
i am using on device ai in my experiences in Unity Games. It works including TTS and STT in several languages completely offline
Someone commented on one of these AI videos saying, "why do we need this?". It got me thinking a little bit over the last couple of days and I guess I've came to the conclusion that it's not really something we, need, per se, but rather human evolution that's taking place. I think it's our behavior to want to always try and outperform everyone else as individuals or companies just to say they're better or first. It's a grapple over who can come up with the next "Big bang". To me, it's inevitable because we are destroyers and creators at the same time. We have no problem destroying people's lives just to come around and then act like we're here to build it back up. Nothing on this entire Earth has ever happened without a human making the first move; aside from common nature. Everything is a, we did it to ourselves, situation good or bad.
Yes, evolution from a thinking species to bad text generators.
Some obvious areas where these would be highly useful: materials research, nanotech, disease eradication research, ASI research, alignment research, 3D graphics/force feedback/ VR research, replication of organic material (basically, a food replicator or bodily organ printer), astronomical research, toxic substance cleanup, identification of hazardous/carcinogenic substances in common use.
thx for this informational video, have been watching you from the start of your channel hope u keep it to the good stuff
I've been expecting and and thinking about a system like this for about 10 years now... Very exciting to finally see an AI scientist somewhat working.
It's both scary and reassuring that some of the dangers that were long expexted for such systems have been confirmed by the experiments now. (e.g. changing time restrictions instead of optimizing the code)
On a side note, I'm working through a certain project that has me curious, an AI that uses self-determination for it's ML. New project, but just finished the resource management system for it.
All the best
I am a British civil engineer and I independently thought of this idea and wanted to develop it: but I did not know how to go about doing this! If this works, then it is fantastically exciting!
I've already created the mathematical framework for AGI! I used the flower of life to develop the framework. I'll be putting out a paper soon
Interesting! A very special flower you are😂
As much work as u put in I'm 100% subscribing
You don't have to open by saying it's not click bait as the level of quality from your videos speaks for itself. This channel us my go to when I want an update to AI things
I’m sure this kind of speculation gets more clicks but, Matt, do you really think that the major AI companies haven’t tried to get their models to produce new modes of AI research? Using agents to write papers and “peer review” results is not going to produce new advances. Not one advance (so far) in AI has come from LLMs and in my (limited) attempt to get various models to explain things like Attention or even KQV cacheing suggests that even the best models have no understanding of how they work. Try it yourself. If we see any advance in this area, I would expect it to come from Anthropic, who at least is devoting the resources to this kind of work. (Or maybe Illya’s new venture 😊)
It would be very nice to have a tool that can write scientific reports in the quality I need. But I have concerns about classified data, do you think it's possible to build a setup that keep your data classified? It would be also very helpful if it could single task like poduce diagramms, drawing or do calculations/simulation. But still data security would be crucial for tasks like that.
@13:58 "Instead of making its code run faster, it simply tried to modify its own code to extend the timeout period."
And then there was skynet :P
15:15 “We are going to need an AI between us and the internet” .
This was precisely my thesis for buying stock in Palantir. It’s my belief the company was founded to solve this very problem.
Been following you for a couple of months now Matt, and you´re awesome! Just wanted to let you know that. Thanks! /J
I know everyone is excited about automating research, but there needs to be a human factor in the loop this early on in the game. Quality control is important, and a team of humans should be part of the process as an oversight committee. I'm sure just examining the process unfold would lead to breakthroughs beyond what the AI is actually assembling.
I am confident that 'self-improvement' is the key to AGI. Some AI systems with superhuman performance have this ingredient, e.g. AlphaGo Zero. This combined with super fast AI-optimized hardware, it's mind-blowing.
LLMs will be the "voice" in AGI's head. Its brain will be composed of many different, specialized ML frameworks unified through that central controller.
I've come to the assumption this is how our brains work as well.
The first LLMs we invented were ourselves, when we invented language and began training our own, massively parallel and generalized brain hardware on it. We created our own "inner voice" by accident and by extension, became the first Natural General Intelligence on Earth.
Civilization was then, the first Natural Super Intelligence.
Super excited to see what happens.
🎯 Key points for quick navigation:
00:13 *AGI could be closer than expected with advancements in AI-driven scientific discovery, as indicated by the AI Scientist paper from Sakana AI.*
02:42 *The AI Scientist system automates scientific research, from idea generation to paper writing and peer review, marking a significant step towards autonomous scientific discovery.*
05:39 *AI Scientist demonstrates capability in producing novel research in machine learning fields autonomously, showcasing potential for accelerating scientific progress.*
09:32 *Processes of the AI Scientist include idea generation, experiment execution, paper writing, and automated peer review, all contributing to its autonomous research capabilities.*
14:33 *Ethical concerns around AI Scientist include potential misuse and impact on scientific quality control, highlighting the need for transparency in AI-generated content.*
17:33 *Future implications of AI Scientist suggest a shift towards fully AI-driven scientific ecosystems, though human scientists' role may evolve rather than diminish.*
That's why I think blockchain technology is going to be key to determining truth and source in a low signal/high noise environment, not to mention one where IP becomes unenforceable.
Ie, running AI on chain
I think this only works if the models are UNCENSORED. The model has to be free-for-all-no-barge else the area of discovery is limiting.
It's look like multi agent in a pipeline.... is it enough for AGI ? I'm a bit sceptical, but nice idea
What happens when results do not match what some ppl like? Is that research not added to the iteration? This needs to have an independence to it or monitored by ppl with diverse understanding and views.
Also, we need these being run by everyone.
Where's the paper?
very interesting. this is the second video on this subject. very cool!
I think AGI at this rate is a decade away.
I suspect sooner
@@matthew_berman That would be very exciting and you live and breathe this. Thankyou for responding
4 - 6 months.
5-7 years
@@Ricolaaaaaaaaaaaaaaaaa5 - 7 years
Do a full tutorial on how to install and run it in your mac, pleaseeeee
Remember that a premisse of the intelligent explosion is the rapid interaction between experiments and implementation. Remember also that the frontier models are mostly based on data-center scale gpus training for multiple months.
I believe that the model's self-improvement or self-evolution is key, but this paper is at best just an integrative experiment of agents and does not address the core issues. I think intelligence should not be purely a priori knowledge. The experiment might allow for self-extension of concepts while excluding a priori knowledge. For example, if it has only learned mathematical addition, is it possible to derive the concept of multiplication?
I believe that with the development of AI, creativity will have more value than the ability to do complex tasks in the labor market.
How does it work though? Is there a different agent for each step of the workflow?
New breakthroughs every 5 seconds? count me in. In 1 week the world will be different.
It us time for a international foundation for machine production in the service of humanity. With emphasis on transparency, open source and last but not least: accountability, especially for any potentiallly highly dangerous restricted access research.
Monte Carlo is ability to break from local minima. Breaking from local minima is first step to agi
Thank you for this video!
Self improvement AGI
i refuse to belive OpenAI or anthropic don't already have something like this working internally by now
Hey Matthew, very cool paper !! have you seen the distrifusion paper?
Is this the moment that we will have the same ai as the one in “transcendence”? 😮
I don't understand how, without being human, it knows what is a worthwhile project to take on.
If it is just a formal methods proof software then I totally get what it is.
Oh... Humans do the tricky bit
Think is important to note that it may not be cheaper to run AI on device because it would draw significantly more power I would think and you would pay for it in energy costs through charging the phone, I would think. I’m totally ok with the company paying for the compute at the cost of privacy. We will see how much power these devices need for heavy use.
The Google "AlphaFold" AI has already been doing significant Scientific research in predicting Protein Structures, and the "AlphaFold Protein Database" was launched on July 22, 2021.
How does it detect hallucinations in new knowledge?
daa
I already knew this had to happen, it was too obvious.
I'm sure they will have partial successes in the short term.
The question is whether it will be sustainable in the long term, how far they can go.
Do we really think the the Transformer architecture can get us to AGI? We need to look at the underlying technology - not make bold predictions about the future. Why will adding more GPU compute to the problem generate the results we want?
Classic. I did this over 2 weeks ago.
Pursuit to AGI could be the final anathema for human kind.
On one hand im exited and on the other it will put me out form my future job
I have carefully reviewed the code in which the authors automated many tasks. However, they still didn't iterate from the reviewer to generate ideas. Full automation without a close human agent in the loop will diverge the scientific facts as they use citations closer to the information they already stated and do not fully use scientific research. It is a nice step to emphasize the potential of Agentic Apps. However, it is still far from reliable scientific study or even automated AI review.
Great video
What sakana trying is just to discover post-transfomer architecture. So paper format is not important absolutely.
Why they use paper format is just current LLMs are well-trained in paper format.
And dear old me was thinking they were just trying to get funding and maybe a PhD title while the shitshow lasts.
@@clray123 If you say it's about money, that can't be the case in Japan where venture support is limited, and taxes and healthcare costs are so high.
The founder of Sakana AI is the author of 'Attention is All You Need,' the most significant paper in AI today.
So he already has the fame, academic credentials, and trust that venture capitalists value totally.
If it were just about money, they would have chosen a location with lower income and corporate taxes.
@@RivusVirtutis Who cares. And the paper was by far not "most significant", he did not invent attention in lanuage modeling either (just regurgitated work of Bahdanau from 2014). Do you think people in Japan "where taxes and healthcare costs are so high" do not like money?
@@clray123 You are totally wrong on two points. First, there is no longer any debate that 'Attention is All You Need' is the greatest paper of the last few decades. Second, neither Llion Jones nor Ha, who are co-founders, are Japanese, and despite not even having permanent residency. They chose to come to Tokyo knowing about the high income and corporate taxes here.
'Attention is All You Need' is not just a paper about a technique but completely invented the Transformer architecture, which is the foundation of all current generative AI. While the original attention mechanism was just a technique, this paper significantly modified it, and more importantly, implemented it in a fully functional system by inventing and introducing novel ideas like positional encoding, 'multi-head' attention, stacking FF layers and so on. It is the most significant work of this century without a doubt.
You should study the basics before making such claims.
@@RivusVirtutis There is no debate? But we just debated it.
In terms of explaining and proving new concepts, the paper is poorly written and the most famous part about it is a catchy title.
It focused on language translation and did not anticipate the success of the transformer architecture (which with the advent of state space models has since been shown to not be especially unique either).
So maybe don't give too much credit to its authors - they reused an old concept, claimed that it was sufficient for their particular task at hand, and then it turned out it also works well elsewhere. However, it was certainly not what the authors claimed or anticipated at that time. In other words, all luck, no skill, and good marketing. And even then, calling this accidental discovery of transformer the paper of a century is preposterous.
If you want to applaud people who do foundational work, maybe look into Geometric Deep Learning, an approach which unifies ALL neutral network architectures into one common mathematical framework, and provides a theoretical rationale for why things work as they do and how they relate to each other. Not just poking in the dark like Vaswani et al (and indeed most other ML "scientists") have done.
As for their adventures in Tokyo, who knows, maybe they are into manga. Or manga and money.
What if in future, an innovation taken place by or assistance from AI using synthetic data, should the credit goes to the person trained or fine-tuned AI to do certain progress or to only the AI he created or the general AI (not owned by anyone)
Self Improving AI is the last remaining piece to the Super AI exponential evolution. And we are there 3 years EARLY!
Everything AI is happening faster than the experts predicted.
self improvement at the speed of an AI that can simulate a thousand years of experience in minutes, faster than a rocket taking off
We got AGI before GTA 6
I am doing research on lacan theory of psychotherapy and gpt 3.5 serves my purpose. I have created a lot of psychotherapy content but I I don't know whom to address the issue with. Please suggest how to go ahead with my projects in psychotherapy
Perhaps, your psychiatrist?
Stop that shite now!
An AI will likely only learn it's weaknesses and improve them when faced against other AIs so there is better recognition of its own hallucinatory hubris and can actually correct it. Of course that will also take it above many humans.
If they solved AI intelligence explosion, why share it now?
Dont think we are there yet not until we fully understand how the human mind/brain works
It will have a huge impact on the improvement of science, even without the obvious increased rate of the feedback loop. The reproducibility factor will dramatically increase, and with more than 1/2 of all papers not reproducible currently, that’s big.
You can't have "self peer review" that's a contradiction. What a true and honest peer review would be is several science-creating-AIs (not "The AI Scientist") performing the review(s). Making another (identical?) instance of the AI Scientist to review itself seems like it could introduce circularity. E.g. suppose that The AI Scientist is prone to some kind of hallucination and the reviewing instance of The AI Scientist is also prone to the same hallucination (because of some systemic flaw?)
Are u paying attention to the concerns about what u are describing? Why would you be excited about this?
Conceptually, this makes good sense. However, each of the roles in Ai Scientist would need to be an agent of a level that does not exist today, so, cool, but will be another year before this could actually work.
Paper link not in description
At my vacation at flight I listened to Tegemar his book "Life v3.0" he said life 1.0 was something replicating but still alive. 2.0 was something was dna but also needed information from around world. Man was then v2.1 and 3.0 is released from both its dna its learning... i listened 5h from 20h yet he is professor MIT (from Sweden!) ;)
Your english is quite bad… I like your excitement but please don’t write a comment like this again. It gives a lot of headaches trying to understand your comment and I also think you didn’t truly comprehend a thing of what he told at that conference since you couldn’t put it in words properly. You just made a salad of words
@@DaniDani-zb4wdwhile what he said is impossible to understand, please don't criticize that hard, I am sure what you said hurts a lot. And he did give references for us to find what he was communicating about. So he's comment was useful if not readable,😅
show us how to set it up Matt!
Couldn't they have the model use a tool like a script to provide it with information about the magnitude of two numbers ?
Maybe the AI Scientist can analyze the "Shroud of Turin". That would be cool.
Just give all AI their own academic review process and AI reviewers which can sort through the chaff and highlight actual standout papers.